Image Annotation Of Ancient Architecture Based On Attention Mechanism And Multilayer Kernel Pseudo-inverse Learning

Ancient Chinese architecture inherits Chinese historical civilization,and the accurate and informative semantics of ancient architecture is significant for reflecting the history and culture of ancient architecture,digital preservation,and the dissemination of Chinese civilization.However,most current research on 2D image annotation of ancient Chinese architecture lacks deep semantic information such as dynasties and regions,making semantic annotation based on images of ancient Chinese architecture a topic of great interest.Meanwhile,in the process of establishing an annotation model,effective image feature extraction and semantic analysis are essential to improve the accuracy and efficiency of the model.The visual attention mechanism dynamically focuses on the primary semantic content of the image to reduce the computational cost of unimportant data.The pseudo-inverse learning algorithm can accelerate the training efficiency of the model.Therefore,to enhance the performance of the image annotation model of ancient architecture and enrich its semantic information,this paper studies the image annotation method of ancient Chinese architecture based on visual attention mechanism and multi-layer kernel pseudo-inverse learning.The main contents are as follows:(1)Aiming at the uniqueness of the roof ridges of ancient Chinese architecture,we present an image annotation method of ancient architecture based on visual attention mechanism and graph convolutional network to effectively reflect the historical information of ancient architecture.First,regions of roof ridge decoration of ancient architecture images are extracted by introducing a visual attention mechanism in the convolutional neural network.Then,the correlation between the ancient architecture labels is further mined with a graph convolutional network,and a classifier is generated to avoid massive incorrect labels in the output.Finally,the classifier is acted on the extracted roof ridge decoration features and is trained for multi-label loss.The experiments on the constructed dataset of ancient architecture demonstrate that this method not only enriches the semantic information such as history and culture of ancient architecture images but also improves the accuracy of the annotation model.(2)To improve the efficiency of image annotation of ancient architecture,an image classification model based on a convolutional neural network and multilayer kernel pseudo-inverse learning is presented to address the problems of gradient disappearance and slow convergence in the image classification method of model training by backpropagation in(1).Firstly,the image features are extracted by selecting the optimal model from a pre-trained convolutional neural network.Secondly,a multilayer kernel pseudo-inverse learning classifier is constructed for model training based on kernel pseudo-inverse learning and kernel pseudo-inverse learning autoencoder.Finally,experiments on MNIST,CIFAR-10,and ancient architecture datasets demonstrate that this method outperforms the CNN-Softmax model in classification accuracy and efficiency.(3)On the basis of(2),we propose an ancient architecture annotation model based on multi-layer kernel pseudo-inverse learning and graph attention network to further improve the training efficiency and model accuracy of the model.First,the convolutional neural network with a multilayer kernel pseudo-inverse learner trains the extracted image patches of roof ridge decoration of ancient architecture for reducing the training time of the model.Then,the feature learning capability of the nodes(ancient architecture labels)is further enhanced by a graph attention network.Finally,the experiments on the ancient Chinese architecture dataset verify that this method can effectively improve the annotation performance and training efficiency of the model.